Mobile AI 2021 Learned Smartphone ISP Challenge Forum

Go back to competition Back to thread list Post in this thread

> TensorFlow version restriction

If the server isn't capable to support models converted using the TF2 converter then it's missing out on a significant amount of improvements and it might impact a lot of participants. Since the TF2 converter supports converting models from TF1 I don't see any issue updating the server to support TFLite models converted using TF2. Not supporting this would restrict the participants to only using TF1 which is absolutely not ideal. I hope you'd agree here.

Posted by: sayak @ Feb. 3, 2021, 6:14 a.m.

Hi Sayak,

Thank you for your feedback.

The TF1.15.0 restriction here is to consider the mobile deployment on the target device.
Any TF version later than 1.15.0 may cause some issues unpredictable.
We hope you understand this.

Thank you.

Posted by: HsienKaiKuo @ Feb. 3, 2021, 6:33 a.m.

I see.

But TF2 converted TFLite models are very well supported on both Android and iOS platforms. So, do you mind shedding a bit more light on this?

Posted by: sayak @ Feb. 3, 2021, 7:05 a.m.

The reason is not about the OS (e.g. Android or iOS).
The main reason for the restriction is because we will run the submitted models on our developed computing platform Dimensity 1000+, which contains the powerful computing processing unit APU.
For more details about Dimensity 1000+, please check:
Currently, TF2 is not fully supported by Dimensity 1000+.
That's why we restrict the TF version to 1.15.0.

Posted by: MHChen @ Feb. 4, 2021, 1:15 a.m.
Post in this thread